What is League of Legends?

• A free-to-play, multiplayer online battle arena (MOBA) videogame
• Developed and published by Riot Games
• ~100 million active players a month
• ~100k viewers during peak hours on Twitch.tv
• Active and popular professional eSports scene
• Numerous 3rd party websites and books on gameplay strategies

How is League of Legends Played?

• Primary mode: Summoner’s Rift
• 5 vs. 5 human players, plus AI minions for each team
• Main objective: destroy enemy’s base (aka The Nexus)
• Sub-objectives: destroy enemy’s defense turrents
• Key objectives: slay dragons and Baron Nashor for boosts

How is League of Legends Played?

• Players select 1 of 141 champions to represent them
• Each champion has unique set of attributes and abilities
• Gain power by accumulating: XP to level up, and gold to buy enhancement items
• Accumulate XP/gold by killing minions, monsters, and enemy players, and taking objectives
• Killed players must wait before respawn

What are Team Roles?

• In standard team composition, each player is designated one of fives roles based on:
• Their champion’s attributes and abilities
• Their intended location on the map near start of a match (around 1:25-1:30) – called the laning phase
• The Roles: Top Laner, Jungler, Mid Laner, Botlane Carry, Support

Top Laner

• Fights in the top lane of the map during laning phase
• Key attributes/responsibilities: front line tankiness, deal moderate-to-high damage, ability to safely initiate team fights

Jungler

• Fights in the jungle areas of the map during laning phase
• Key attributes/responsibilities: ability to efficiently kill most jungle monsters during laning phase, provide ganks for in-lane teammates, place wards on map
• Can deal high burst damage to neutral monsters via Smite

Mid Laner

• Fights in the middle lane of the map during laning phase
• Key attribute/responsibility: deal high burst, ability power-based damage
• Champion is usually a mage or an assassin

Botlane Carry

• Fights in the bottom lane of the map during laning phase
• Key attributes/responsibilities: back line squishiness, deal high sustained attack damage, be the highest damage dealer in match
• Champion is usually ranged (e.g., guns, arrows, etc.)

Support

• Fights/assists in the bottom lane of the map during laning phase
• Key attributes/responsibilities: help Botlane Carry survive and achieve kills, provide healing, shielding, and/or other utilities during team fights, place wards on map

The Problem

• Humanly impossible to visually observe every match and record team roles manually.
• Can we use match data and machine learning to distinguish all five team roles from each other?

Team Roles – Match Overview, eSports

Assumed Team Roles – eSports Match

Explicit Team Roles – Draft Pick Mode Selection Screen

Team Roles – Match Overview, Blind Pick Format

Team Roles – NA LCS 2018 Spring Split, All Matches

Why Should We Care?

• Most LoL analytics websites only track Draft Pick mode matches because of extracting explicit team roles
• Some only track Ranked (ladder) matches
• Normal Draft Pick requires owning/having available at least 20 champions
• Ranked Draft Pick requires owning at least 20 champions plus being Gamer/Summoner Level 30
• Granting sites ability to analyze team role performace in Blind Pick matches adds another avenue of engagement for newcomers/casuals in order to graduate to dedicated/hardcore
• Provide additional team role-focused metrics when testing new/reworked champions

Data Overview

• Data for all matches in the 2018 Spring Split seasons of the North America LoL Championship Series (NALCS), the Europe LoL Championship Series (EULCS), LoL Champions Korea (LCK), and LoL Master Series (LMS), plus the 2018 Mid-Season Invitational (MSI)
• Each match data accessed using API domain https://acs.leagueoflegends.com/v1/stats/game/

Synopsis of Individual Match Player Datasets

• NA LCS: 117 matches, 1170 observations
• EU LCS: 115 matches, 1150 observations
• LCK: 246 matches, 2460 observations
• LMS: 144 matches, 1440 observations
• MSI: 74 matches, 740 observations

Synopsis of Per-Match Average Datasets

• NA LCS: 56 players
• EU LCS: 57 players
• LCK: 81 players
• LMS: 69 players
• MSI: 74 players
• (More technical details below…)

Obtaining API and Match IDs

Sample Collection of Match IDs

Cleaning Each Match for Player Stats

• JSON response object contains two sub-objects (participants and participantIdentities) that contain indvidual player stats
• participantIdentities contains 10 players’ usernames under sub-object player
• participants contains two more sub-objects (stats and timeline)
• basically, participants is flattened out, then joined with player
• final product for single match is 10 rows and nearly 150 columns/variables
• process repeated for all other matches, and tables are concatenated together

Sample of MSI 2018 Match Player Dataset

Creating Player Per-Match Average Datasets

• Observations from the match player datasets are grouped together by team, player name, and then assumed team role
• Each group’s variables’ statistics are calculated to mean (average) values.

Sample of MSI 2018 Player Per-Match Average Dataset

Box Plots 1

Box Plots 2

Box Plots 3

Box Plots 4

Results: Training the Model (Per-Game Averages Data)

train_fit.km <- kmeans(train_avgs_data_scaled, 5, iter.max = 1000)

Results: Validating the Model (Per-Game Averages Data)

knn_pred_test_avgs <- get.knnx(train_fit.km$centers, 
                               test_avgs_data_scaled, 1)$nn.index[, 1]

Results: Testing the Model (Individual Matches Data)

knn_pred_test_single_games <- get.knnx(train_fit.km$centers,
                                       all_leagues_match_player_stats_scaled,
                                       k=1)$nn.index[, 1]

Confusion Matrix

## Confusion Matrix and Statistics
## 
##                           
## knn_pred_test_single_games BOTCARRY JUNGLE  MID SUPPORT  TOP
##                   BOTCARRY     1321      1   10       0   68
##                   JUNGLE          0   1378    1      14    4
##                   MID            19      1 1321       0  120
##                   SUPPORT         0      7    1    1376    0
##                   TOP            52      5   59       2 1200
## 
## Overall Statistics
##                                           
##                Accuracy : 0.9477          
##                  95% CI : (0.9422, 0.9528)
##     No Information Rate : 0.2             
##     P-Value [Acc > NIR] : < 2.2e-16       
##                                           
##                   Kappa : 0.9346          
##  Mcnemar's Test P-Value : NA              
## 
## Statistics by Class:
## 
##                      Class: BOTCARRY Class: JUNGLE Class: MID
## Sensitivity                   0.9490        0.9899     0.9490
## Specificity                   0.9858        0.9966     0.9749
## Pos Pred Value                0.9436        0.9864     0.9042
## Neg Pred Value                0.9872        0.9975     0.9871
## Prevalence                    0.2000        0.2000     0.2000
## Detection Rate                0.1898        0.1980     0.1898
## Detection Prevalence          0.2011        0.2007     0.2099
## Balanced Accuracy             0.9674        0.9933     0.9619
##                      Class: SUPPORT Class: TOP
## Sensitivity                  0.9885     0.8621
## Specificity                  0.9986     0.9788
## Pos Pred Value               0.9942     0.9105
## Neg Pred Value               0.9971     0.9660
## Prevalence                   0.2000     0.2000
## Detection Rate               0.1977     0.1724
## Detection Prevalence         0.1989     0.1894
## Balanced Accuracy            0.9935     0.9204

Current Classification System

• Next: machine learning approach (technical)

Feature Set for Machine Learning 1

Feature Set for Machine Learning 2

Feature Set for Machine Learning 3

Machine Learning Approach

1. Put together all five datasets for player per-match averages, then split the entire set into two subsets: one for training and one for testing.
2. Create/train a model using the k-means clustering algorithm on the training set. Since we already know how many team roles exist, we set k=5.
3. Test/validate the model using the k-nearest-neighbor (knn) classification algorithm on the testing set. Since there are only five possible neighbors, and we want to classify each observation to exactly one group (team role), we set k=1.

Machine Learning Approach (continued)

4. Put together all five datasets for single match player stats.
5. Test the model using knn on the new dataset.
6. Evaluate performance of Steps 2, 3, and 5 using a confusion matrix.

Note: features are scaled using z-score.

Summary

• 21 out of ~150 variables were chosen for model’s feature set
• k-means clustering on per-match averages training set was 100% accurate
• Used centroids produced from clustering as the model for validation and testing sets
• Model used for knn classification on per-match averages validation set was 100% accurate
• Used same model on “real world” single match performances test set
• knn classification on test set was 94.7% accurate

Caveats / Limitations

• Sample size vs. population
• Patches ==> possible meta changes
• Normal / Blind Pick / vs. AI modes are lower stakes
• Unique champions (e.g. Vladimir, Swain, Gangplank, Ivern)
• Varying levels of competition

Application of Solution

• Train model by running k-means on sample of Draft Pick mode matches, then run knn on subsequent matches of any game mode to classify team roles
• When creating new champions or reworking current ones, while in test, can run the model to see if champion is performing certain team roles as expected

Future Refinements

• Larger datasets: virtually endless number of non-custom matches can be accessed off of one username alone due to recursive nature of accessing match history
• Additional data (i.e., the timeline API using match IDs)
• Alternative k-means training
• Creating models per division
• Creating models per patch

Final Thoughts

• Draft Pick game modes allow safe assumptions on team roles
• Limits the number of matches analytics sites can use for analyzing champion and team role performances
• While not absolutely perfect, model could provide excellent tool for classifying team roles in Blind Pick Summoner’s Rift game modes,where team roles cannot be assumed based on Participant ID alone

The End